workflow-filesystem-offload
Skill1 starsChanged 6 months agoMITView on GitHub
When a tool returns large output (a research synthesis, a big file read, long logs, a wide search result), write it to a scratch file and keep only a ~100-token summary + the file path in context — then read back specific line ranges with grep on demand instead of holding the whole blob. Trigger whenever a single tool result is more than ~50 lines / a few KB.
In this repository33 other files it gives its agents.
- skill-process-gotchasSkill
- workflow-cache-disciplineSkill
- workflow-codegraphSkill
- workflow-conductorSkill
- workflow-context-condenserSkill
- workflow-context-hygieneSkill
- workflow-criticSkill
- workflow-deadline-disciplineSkill
- workflow-deep-researchSkill
- workflow-done-definitionSkill
- workflow-edit-formatSkill
- workflow-effort-routingSkill
- workflow-escalateSkill
- workflow-execute-from-planSkill
- workflow-learn-frameworkSkill
- workflow-long-running-processesSkill
- workflow-lspSkill
- workflow-memory-curationSkill
- workflow-plan-contractSkill
- workflow-plan-firstSkill
- workflow-planSkill
- workflow-quality-barSkill
- workflow-repomapSkill
- workflow-retrieve-before-actSkill
- workflow-spec-drivenSkill
- workflow-stuck-detect-resetSkill
- workflow-stuckSkill
- workflow-subagent-isolationSkill
- workflow-subtask-loopSkill
- workflow-task-finishSkill
- workflow-task-startSkill
- workflow-tool-selectionSkill
- workflow-verify-enhancedSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 2 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 21
When a tool returns a large result, write it to .hermes-scratch/<descriptive-name>.md in the project (create the dir if missing.
- Very specific: line 23
Keep in context only: a one-line summary + the path + the line count.
- Distinctive: line 25
Read back only what you need, on demand:
- Very specific: line 27
read the specific line range (e.g.
- Distinctive: line 28
Never paste the whole blob back into context to "remember" it — the file IS the memory.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
workflow-filesystem-offload
<!-- TRIGGERS WHEN: a single tool result is large (>~50 lines / a few KB) -->Holding large tool outputs in context burns the window and degrades coherence (the same pressure the condenser relieves after the fact — this prevents it up front). Offload, don't hold.
The rule (deterministic)
- When a tool returns a large result, write it to
.hermes-scratch/<descriptive-name>.mdin the project (create the dir if missing; it's git-ignorable scratch). - Keep in context only: a one-line summary + the path + the line count.
e.g.
→ deep_research on BLAKE3 saved to .hermes-scratch/blake3-research.md (412 lines): spec + 8 sources + test vectors. - Read back only what you need, on demand:
grep -n "<term>" .hermes-scratch/<file>to locate, then- read the specific line range (e.g.
sed -n '120,160p'or the file-read tool with offset/limit).
- Never paste the whole blob back into context to "remember" it — the file IS the memory; re-read the slice you need.
When to use
deep_research/synthesizereports → offload, keep the gap_note + path.- Large file reads you only need a section of → note the path, grep the symbol.
- Long command logs / test output → save, keep the failing lines only.
This pairs with [[workflow-context-condenser]] (offload prevents the bloat; condense relieves it once it's already there) and [[workflow-cache-discipline]] (keeping context small + stable also protects the prompt cache).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.